Ashwini T (@AshwinTRIP786) header
Ashwini T (@AshwinTRIP786) profile picture

Hello Software

India
Joined August 2023
Day 17/61 — What Is Model Distillation? Today I learned how a large AI model can transfer its knowledge to a smaller one. The goal is simple: make models faster, cheaper and easier to run without losing too much of their capabilities. and...
1
2
16
3/4 I also learned about something called "dark knowledge." For example, after "She poured milk into her..." Cereal might be the most likely answer, but coffee is also reasonable. Soft labels help the student learn these relationships between possible answers.
1
1
3
4/4 Why does this matter? Smaller models need less memory and compute, reducing cost and latency. The trade-off is that they may lose some capabilities. My takeaway: It's not just about making models smaller, but preserving as much useful knowledge as possible.
1
3
Quick SQL test. Which query returns duplicate email addresses? A. GROUP BY email B. SELECT DISTINCT email C. GROUP BY email HAVING COUNT(*) > 1 D. ORDER BY email Let's see how many get this right.
1
38
Day 16/61 — MCP vs. gRPC Today I learned how MCP and gRPC help AI agents communicate with external services. Both solve the problem of connecting systems, but they approach it differently. Understanding when to use each one was the interesting part. and....
2
3
78
3/4 One key difference I learned: MCP provides tools with descriptions that help AI understand what they do and when to use them. gRPC provides service methods and definitions, but usually needs an adapter layer to translate AI intent into RPC calls.
1
1
10
4/4 My biggest takeaway: MCP helps AI agents discover and use tools. gRPC handles fast, efficient service communication. They aren't necessarily competitors. MCP can handle AI-facing interactions while gRPC powers communication behind the scenes.
1
11
Same AI. Same subscription. Different outcomes. One person asks AI questions. Another delegates tasks, builds workflows, and runs multiple agents. The difference isn't access to AI. It's how you use it. Are we underestimating AI?
1
3
55
Day 15/61 — LLM Observability Today I learned why monitoring AI agents is different from monitoring traditional applications. An API can return 200 OK, but the agent can still give a completely wrong answer. That's where LLM observability comes in. and....
3
3
204
3/4 Another interesting concept: LLM-as-a-Judge. A second model evaluates an agent's output based on relevance, safety, tool selection and efficiency. MLflow also helps manage prompt versions and compare evaluations.
1
1
22
4/4 Some production practices I learned: • Use a proper tracking database. • Log traces asynchronously. • Capture errors and sample successful traces. • Run evaluations in CI. My takeaway: Building AI agents is important, but understanding their behaviour matters too.
1
10
Day 14/61 — NLP vs. NLU vs. NLG Today's learning was about three terms that are closely connected but serve different purposes: NLP, NLU and NLG. I used to see these terms together, but now I understand how they fit together. and..
1
2
95
3/4 And NLG (Natural Language Generation). This is about making computers generate human-like language from data or input. It involves three stages: • Text planning • Sentence planning • Realization (correct grammar and wording)
1
1
1
4/4 The simplest way I understood the difference: NLP = Working with human language. NLU = Understanding human language. NLG = Generating human language. My biggest takeaway: NLU and NLG are both important parts of the broader NLP field.
1
2